opencv-python读写视频

读视频，提取帧

接口函数：cv2.VideoCapture()

通过video_capture = cv2.VideoCapture(video_path)可以获取读取视频的句柄。而后再通过flag, frame = video_capture.read()可以读取当前帧，flag表示读取是否成功，读取成功后，句柄会自动移动到下一帧的位置。读取结束后使用video_capture.release()释放句柄。

一个简单的逐帧读取的程序如下：

import cv2

video_capture = cv2.VideoCapture(video_path)
while True:
    flag, frame = video_capture.read()
    if not flag:
        break
    # do something with frame
video_capture.release()

获取视频信息

为了能更好更灵活地了解并读取视频，我们有时候需要获取视频的一些信息，比如帧率，总帧数等等。获取这些信息的方法是调用video_capture.get(PROP_ID)方法，其中PROP_ID是OpenCV定义的一些常量。

常用的信息及示例如下：

import cv2

video_path = r'D:\peppa\Muddy_Puddles.mp4'
video_capture = cv2.VideoCapture(video_path)

frame_num = video_capture.get(cv2.CAP_PROP_FRAME_COUNT) # ==> 总帧数
fps = video_capture.get(cv2.CAP_PROP_FPS)               # ==> 帧率
width = video_capture.get(cv2.CAP_PROP_FRAME_WIDTH)     # ==> 视频宽度
height = video_capture.get(cv2.CAP_PROP_FRAME_HEIGHT)   # ==> 视频高度
pos = video_capture.get(cv2.CAP_PROP_POS_FRAMES)        # ==> 句柄位置

video_capture.set(cv2.CAP_PROP_POS_FRAMES, 1000)        # ==> 设置句柄位置
pos = video_capture.get(cv2.CAP_PROP_POS_FRAMES)        # ==> 此时 pos = 1000.0

video_capture.release()

句柄位置指的是下一次调用read()方法读取到的帧号，帧号索引从0开始。

使用set(cv2.CAP_PROP_POS_FRAMES)读取指定帧

从上面代码中可以看到我们使用了set方法来设置句柄的位置，这个功能在读取指定帧时很有用，这样我们不必非要使用read()遍历到指定位置。

但问题来了，这种方式读取到的内容和read()遍历读取到的内容是否完全相同？

做个简单的实验，下面用两种方法分别读取同一个视频的[100, 200)帧，然后检查读取的内容是否完全相同，结果是True。

import cv2
import numpy as np

video_path = r'D:\peppa\Muddy_Puddles.mp4'
video_capture = cv2.VideoCapture(video_path)
cnt = -1
frames1 = []
while True:
    cnt += 1
    flag, frame = video_capture.read()
    assert flag
    if 100 <= cnt < 200:
        frames1.append(frame)
    if cnt >= 200:
        break
video_capture.release()

video_capture = cv2.VideoCapture(video_path)
frames2 = []
for i in range(100, 200):
    video_capture.set(cv2.CAP_PROP_POS_FRAMES, i)
    flag, frame = video_capture.read()
    assert flag
    frames2.append(frame)
video_capture.release()

frames1 = np.array(frames1)
frames2 = np.array(frames2)
print(np.all(frames1 == frames2))  # ==> check whether frames1 is same as frames2, result is True

接下来看看利用set读取的效率。还是利用小猪佩奇第一集做实验，这个视频共7788帧，下面分别用两种方法遍历读取视频中所有帧。第二种方法明显比第一种慢得多，所以这就很苦逼了。。。如果帧间隔比较小的话，单纯用read()进行遍历效率高；如果帧间隔比较大的话，用set()设置位置，然后read()读取效率高。
（如果给第二种方法加个判断，每隔n帧读取一次，那么效率确实会提高n倍，可以自行尝试）

import cv2
import numpy as np
import time

video_path = r'D:\peppa\Muddy_Puddles.mp4'
video_capture = cv2.VideoCapture(video_path)
t0 = time.time()
while True:
    flag, frame = video_capture.read()
    if not flag:
        break
t1 = time.time()
video_capture.release()

video_capture = cv2.VideoCapture(video_path)
t2 = time.time()
frame_num = int(video_capture.get(cv2.CAP_PROP_FRAME_COUNT))
for i in range(frame_num):
    video_capture.set(cv2.CAP_PROP_POS_FRAMES, i)
    flag, frame = video_capture.read()
    assert flag
t3 = time.time()
video_capture.release()

print(t1 - t0)  # ==> 76.3 s
print(t3 - t2)  # ==> 345.1 s

读取函数（重点）

上面我们使用了两种方法读取视频帧，第一种是使用read()进行暴力遍历，第二种是使用set()设置帧号，再使用read()读取。两种方法读取到的结果完全一样，但是效率在不同的情况下各有优势，所以为了最大化发挥两者的优势，在写读取帧函数时，就要把两种方式都写进去，由参数来决定使用哪种模式，这样用户可以针对电脑的硬件做一些简单实验后自行决定。

# -*- coding: utf-8 -*-
import os
import cv2


def _extract_frame_mode_1(video_capture, frame_list, root_folder, ext='png'):
    """
    extract video frames and save them to disk. this method will go through all
    the frames using video_capture.read()

    Parameters:
    -----------
    video_capture: obtained by cv2.VideoCapture()
    frame_list: list
        list of frame numbers
    root_folder: str
        root folder to save frames
    ext: str
        extension of filename
    """
    frame_list = sorted(frame_list)
    video_capture.set(cv2.CAP_PROP_POS_FRAMES, 0)
    cnt = -1
    index = 0
    while True:
        cnt += 1
        flag, frame = video_capture.read()
        if not flag:
            break
        if cnt == frame_list[index]:
            filename = os.path.join(root_folder, str(cnt) + '.' + ext)
            cv2.imwrite(filename, frame)
            index += 1


def _extract_frame_mode_2(video_capture, frame_list, root_folder, ext='png'):
    """
        extract video frames and save them to disk. this method will use
        video_capture.set() to locate the frame position and then use
        video_capture.read() to read

        Parameters:
        -----------
        video_capture: obtained by cv2.VideoCapture()
        frame_list: list
            list of frame numbers
        root_folder: str
            root folder to save frames
        ext: str
            extension of image filename
        """
    for i in frame_list:
        video_capture.set(cv2.CAP_PROP_POS_FRAMES, i)
        flag, frame = video_capture.read()
        assert flag
        filename = os.path.join(root_folder, str(i) + '.' + ext)
        cv2.imwrite(filename, frame)


def extract_frame(video_path, increment=None, frame_list=None,
                  mode=1, ext='png'):
    """
    extract video frames and save them to disk. the root folder to save frames
    is same as video_path (without extension)
    
    Parameters:
    -----------
    video_path: str
        video path
    increment: int of 'fps'
        increment of frame indexes
    frame_list: list
        list of frame numbers
    mode: int, 1 or 2
        1: go through all the frames using video_capture.read()
        2: use video_capture.set() to locate the frame position and then use
        video_capture.read() to read
    ext: str
        extension of image filename
    """
    video_capture = cv2.VideoCapture(video_path)
    frame_num = int(video_capture.get(cv2.CAP_PROP_FRAME_COUNT))

    if increment is None:
        increment = 1
    elif increment == 'fps':
        fps = video_capture.get(cv2.CAP_PROP_FPS)
        increment = round(fps)

    if frame_list is None:
        frame_list = [i for i in range(0, frame_num, increment)]

    if frame_num // len(frame_list) > 5 and mode == 1:
        print("the frames to be extracted is too sparse, "
              "please consider setting mode = 2 to accelerate")

    root_folder = os.path.splitext(video_path)[0]
    os.makedirs(root_folder, exist_ok=True)
    if mode == 1:
        _extract_frame_mode_1(video_capture, frame_list, root_folder, ext)
    elif mode == 2:
        _extract_frame_mode_2(video_capture, frame_list, root_folder, ext)
    video_capture.release()


if __name__ == '__main__':
    video_path = r'D:\peppa\Muddy_Puddles.mp4'
    extract_frame(video_path, increment=30, mode=2)

将图像写为视频

写视频没有那么多需要注意的地方，主要使用的接口函数是cv2.VideoWriter(video_path, fourcc, fps, size)，该函数的主要注意点是入参的设置，video_path是输出视频的文件名，fps是帧率，size是视频的宽高，待写入视频的图像的尺寸必需与size一致。其中不太容易理解的是与视频编码相关的fourcc，该参数的设置需要使用另外一个接口函数：cv2.VideoWriter_fourcc(c1, c2, c3, c4)，c1-c4分别是四个字符。

示例

因为获取图像的方式多种多样，而写视频又比较简单，所以不太适合将这部分写成函数，下面以一个例子呈现。

video_path = r'D:\peppa\Muddy_Puddles.avi'
root_folder = r'D:\peppa\Muddy_Puddles'

fourcc = cv2.VideoWriter_fourcc('X', 'V', 'I', 'D')
fps = 25
size = (1920, 1080)

video_writer = cv2.VideoWriter(video_path, fourcc, fps, size)
for i in range(0, 7788, 30):
    filename = os.path.join(root_folder, str(i) + '.png')
    image = cv2.imread(filename)
    video_writer.write(image)
video_writer.release()

fourcc

fourcc有时候需要多尝试一下，因为不同电脑里安装的编解码器可能不太一样，不见得随便设置一个参数就一定能成功，fourcc有非常多，比如：

paramters	codec	extension
(‘P’,‘I’,‘M’,‘1’)	MPEG-1	avi
(‘M’,‘J’,‘P’,‘G’)	motion-jpeg	mp4
(‘M’,‘P’,‘4’,‘V’)	MPEG-4	mp4
(‘X’,‘2’,‘6’,‘4’)	H.264	mp4
(‘M’, ‘P’, ‘4’, ‘2’)	MPEG-4.2
(‘D’, ‘I’, ‘V’, ‘3’)	MPEG-4.3
(‘D’, ‘I’, ‘V’, ‘X’)	MPEG-4	avi
(‘U’, ‘2’, ‘6’, ‘3’)	H263
(‘I’, ‘2’, ‘6’, ‘3’)	H263I
(‘F’, ‘L’, ‘V’, ‘1’)	FLV1	flv
(‘X’,‘V’,‘I’,‘D’)	MPEG-4	avi
(‘I’,‘4’,‘2’,‘0’)	YUV	avi

上表中的后缀名似乎并不需要严格遵守。

fourcc的参考资料：
https://www.fourcc.org/codecs.php （这个比较完整）
https://blog.csdn.net/tionsu/article/details/81356006
https://blog.csdn.net/weixin_43609992/article/details/106782527
https://www.cnblogs.com/zhangzhihui/p/12503152.html